Early, intermediate and late fusion strategies for robust deep learning-based multimodal action recognition.